What is the best NARX Network training data?

I am using a Non-Linear Autoregressive Neural Network with Exogenous inputs (NARX) to model the way the foot angle reacts to FES pulses. The model is meant to output the pulsewidth value that will generate the desired foot angle value.

I trained the neural network with the signals in the Figure. The foot angles were the input values and the pulsewidth was the output value. I did not obtain good results (negative goodness of fit values) and I think it is because the data is non-causal, seeing as the foot angle has a delay compared to the pulsewidth. If I invert the data (input the pulsewidth and output the angle) the network trains well and has great results, but for control purposes I need the inverse model.

Do you suppose the problem is in fact the type of data used or would you suggest any alteration to the network itself? I have tried to find information about the appropriate type of data to train a NARX network, but couldn't find anything.